DRIDNet: Diffusion-Regularized Intrinsic Decomposition Network for Shadow Robust Hyperspectral and LiDAR Classification

Abstract

Multimodal remote sensing data, especially hyperspectral image (HSI) and light detection and ranging (LiDAR) data, have substantially advanced land-cover classification. However, cloud shadows and structure-induced occlusions distort observed spectra and also weaken cross-modal consistency, which degrades classification performance in complex scenes. Most existing HSI–LiDAR classification methods learn directly from observed radiance by using convolutional or transformer backbones, while paying limited attention to the physical image-formation process and illumination-related variability. To address this issue, we propose a diffusion-regularized intrinsic decomposition network (DRIDNet) for shadow-robust HSI and LiDAR classification. At the physical level, a coordinate implicit intrinsic decomposition module (CIIDM) decomposes radiance into an intrinsic reflectance component and a shading-related environmental component, and a spectral denoising diffusion probabilistic model (DDPM) prior regularizes reflectance spectra toward a plausible material manifold, thereby suppressing noise-driven and nonphysical solutions in shadowed regions. At the representation level, a dual-stream pyramid encoder extracts multiscale spectral–spatial cues from HSI and structural cues from LiDAR, while a cross-modal transport routing module (CTRM) performs content-adaptive alignment and routing on high-level scale-aligned multimodal features. Experiments on four public HSI–LiDAR datasets demonstrate that DRIDNet consistently improves over representative baselines. In addition, a dedicated shadow-region evaluation on Houston2013 shows that the proposed method remains more robust than competing approaches, with overall accuracy (OA) gains of 0.65 percentage points on the full shadow region and 14.71 percentage points on a representative hard local shadow patch.